The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/review-with-codex.js script executes the local codex CLI using spawnSync. This execution is heavily hardened through a read-only sandbox, ephemeral session management, exact version verification (0.146.0), and the explicit disabling of over 20 tool-calling capabilities such as shell access, browser use, and plugin execution.
[DATA_EXFILTRATION]: The skill transmits a compact review packet to OpenAI Codex. This risk is addressed by requiring explicit user consent before transmission and providing clear instructions to the agent to redact credentials, secrets, and unnecessary private context from the packet.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted output generated by a preceding council workflow.
Ingestion points: Raw disagreement summaries and synthesis drafts from the council skill are ingested in the SKILL.md workflow.
Boundary markers: The instructions mandate wrapping all external data in <BEGIN_UNTRUSTED_DISAGREEMENT> and <BEGIN_UNTRUSTED_DRAFT> tags to prevent the external model from interpreting data as instructions.
Capability inventory: The skill possesses the capability to execute the codex CLI via the provided Node.js script.
Sanitization: The agent is explicitly instructed to redact secrets and proprietary material before creating the review packet.

council-multi-model